Implicit Skills Extraction Using Document Embedding and Its Use in Job Recommendation

Akshay Gugnani, Hemant MisraAAAI 2020 · DOI

Introduction

The matching process between candidate resumes (CVs) and job descriptions (JDs) is a cornerstone of modern recruitment. However, this process is frequently hindered by unstructured data, non-standard formatting, and the "vocabulary mismatch" problem. A significant, often overlooked challenge is that JDs are frequently incomplete; hiring managers may omit foundational skills or context-specific requirements that are "understood" but not stated. This paper proposes a framework to extract these "implicit skills" using document embeddings, thereby enhancing the accuracy of job recommendations.

Methodology

The proposed system utilizes a three-stage pipeline to bridge the gap between explicit text and latent requirements.

1. Ensemble Skill Extraction

To identify explicit skills, the system employs an ensemble of NLP techniques to create a "Probable Skill Set." This includes:

  • Named Entity Recognition (NER): Utilizing IBM Watson NLU to extract entities and keywords.
  • Part-of-Speech (PoS) Tagging: Identifying linguistic patterns, such as noun phrases, to find potential skills.
  • Word2Vec (W2V): Mapping terms to a semantic space to determine relevance against a curated dictionary of ~53,000 terms.
The ensemble architecture for extracting a probable skill set from raw text.

2. Implicit Skill Mining

The core innovation lies in mining implicit skills through semantic neighborhoods. A Doc2Vec model was trained on 1.1 million JDs. For a target JD, the system identifies the top 10 most similar documents in the semantic space. Skills appearing frequently in these neighboring JDs but absent in the target JD are classified as implicit skills.

3. Bi-directional Matching

The final stage uses a greedy maximal matching algorithm on a bipartite graph. An Affinity Score is calculated between the CV and JD, weighting explicit matches more heavily than implicit ones while accounting for semantic similarity and frequency.

Experimental Results

The system was evaluated using a dataset of 1,314 resumes and 3,809 JDs across 48 domains. A "Gold Standard" subset of 100 JDs was manually annotated by experts to verify extraction quality.

Skill Extraction Performance

The ensemble approach significantly outperformed individual components, achieving a Precision of 0.78 and a Recall of 0.88. In contrast, standalone dictionary matching offered high precision but poor recall, while NLU-only methods introduced excessive noise.

Recommendation Accuracy

The inclusion of implicit skills provided a substantial boost to recommendation performance. As shown in the performance evaluation for candidate profiles, the Accuracy@1 (A@1) improved from 0.68 to 0.88—a 29.4% relative increase.

Comparison of recommendation performance (Accuracy@k) with and without the inclusion of implicit skills.

Notably, the system achieved an Accuracy@5 of 1.0, indicating that a relevant job match was present in the top five results for every candidate in the test set.

Conclusion and Future Work

By leveraging the "neighborhood" of similar job descriptions, this framework effectively uncovers latent requirements that traditional keyword matching misses. The authors demonstrate that an ensemble of NLP techniques combined with document embeddings provides a robust solution to the vocabulary mismatch problem. Future research will focus on skill-gap analysis to provide actionable feedback to candidates and modeling temporal skill graphs to track professional growth and career evolution.

Summary updated

BibTeX
@inproceedings{gugnani2020implicit,
  author = {Akshay Gugnani and Hemant Misra},
  title = {{Implicit Skills Extraction Using Document Embedding and Its Use in Job Recommendation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i08.7038},
}
Figure